In this thesis, we focus on the problem of image classification and investigate how its performance can be systematically improved. Improving the performance of different computer vision methods has been the subject of many studies. While different studies take different approaches to achieve this improvement, in this thesis we address this problem by investigating the relevance of the statistics collected from the image. We propose a framework for gradually improving the quality of an already existing image descriptor. In our studies, we employ a descriptor which is composed the response of a series of discriminative components for summarizing each image. As we will show, this descriptor has an ideal form in which all categories become linearly separable. While, reaching this form is not possible, we will argue how by replacing a small fraction of these components, it is possible to obtain a descriptor which is, on average, closer to this ideal form. To do so, we initially identify which components do not contribute to the quality of the descriptor and replace them with more robust components. As we will show, this replacement has a positive effect on the quality of the descriptor. While there are many ways of obtaining more robust components, we introduce a joint feature selection problem to obtain image features that retains class discriminative properties while simultaneously generalising between within class variations. Our approach is based on the concept of a joint feature where several small features are combined in a spatial structure. The proposed framework automatically learns the structure of the joint constellations in a class dependent manner improving the generalisation and discrimination capabilities of the local descriptor while still retaining a low-dimensional representations. The joint feature selection problem discussed in this thesis belongs to a specific class of latent variable models that assumes each labeled sample is associated with a set of different features, with no prior knowledge of which feature is the most relevant feature to be used. Deformable-Part Models (DPM) can be seen as good examples of such models. These models are usually considered to be expensive to train and very sensitive to the initialization. Here, we focus on the learning of such models by introducing a topological framework and show how it is possible to both reduce the learning complexity and produce more robust decision boundaries. We will also argue how our framework can be used for producing robust decision boundaries without exploiting the dataset bias or relying on accurate annotations. To examine the hypothesis of this thesis, we evaluate different parts of our framework on several challenging datasets and demonstrate how our framework is capable of gradually improving the performance of image classification by collecting more robust statistics from the image and improving the quality of the descriptor. / <p>QC 20140506</p>

As the availability of unstructured data on the web continues to increase, it is becoming increasingly necessary to develop machine learning methods that rely less on human annotated training data. In this thesis, we present methods for learning from weakly labeled data. We present a unifying framework to understand weakly labeled data in terms of bias and noise and identify methods that are well suited to learning from certain types of weak labels. To compensate for the tremendous sizes of weakly labeled datasets, we leverage computationally efficient and statistically consistent spectral methods. Using these methods, we present results from four diverse, real-world applications coupled with a unifying simulation environment. This allows us to make general observations that would not be apparent when examining any one application on its own. These contributions allow us to significantly improve prediction when labeled data is available, and they also make learning tractable when the cost of acquiring annotated data is prohibitively high.

This June one of the biggest and most popular sports tournaments will take place in Austria and Switzerland, the European soccer championship 2008 (UEFA EURO 2008). Therefore millions of soccer fans in Europe and throughout the world are asking themselves: "Who is going to win the EURO 2008?" Many people, including sports experts and former players, give their guesses and expectations in the media, but there is also a group with financial incentives, like some economists who expect economical increases for the country of the winning team and bookmakers and their customers who directly make money with their beliefs. Some predictions are only guesses, but other predictions are based on quantitative methods, such as the studies of UBS Wealth Management Research Switzerland and the Raiffeisen Zentralbank. In this report we will introduce a new method for predicting the winner. Whereas other prediction methods are based on historical data, e.g., the Elo rating, or the FIFA/Coca Cola World rating, our method is based on current expectations, the bookmakers odds for winning the championship. In particular we use the odds for winning the championship for each of the 16 teams of 45 international bookmakers. By interpreting these odds as rating of the expected strength of the teams by the bookmakers, we derive a consensus rating by modelling the log-odds using a random-effects model with a team-specific random effect and a bookmaker-specific fixed effect. The consensus rating of a team can be used as an estimator for the unknown "true" strength of a team. Our method predicts team Germany with a probability of about 18.7% as the EURO 2008 winner. We predict also that the teams playing the final will be Germany and Spain with a probability of 13.9%, where Germany will win with a probability of 55%. In our study, Italy, the favorite according to the current FIFA/Coca Cola World ranking and Elo ranking, has a much lower probability than these teams to win the tournament: only 10.6%. The defending champion Greece has low chances to win the title again: about 3.4%. Furthermore, the expected performance of the host countries, Austria and Switzerland, is much better in the bookmakers consensus than in the retrospective Elo and FIFA/Coca Cola World ratings, i.e., indicating an (expected) home court advantage. Despite the associated increase in the winning probabilities, both teams have rather poor chances to win the tournament with probabilities of 1.3% and 4.0%, respectively. In a group effect study we investigate how much the classification into the four groups (A-D) affects the chance for a team to win the championship. / Series: Research Report Series / Department of Statistics and Mathematics

La mayoría de problemas en Visión por computador no contienen una relación directa entre el estímulo que proviene de sensores de tipo genérico y su correspondiente categoría perceptual. Este tipo de conexión requiere de una tarea de aprendizaje compleja. De hecho, las formas básicas de energía, y sus posibles combinaciones, son un número reducido en comparación a las infinitas categorías perceptuales correspondientes a objetos, acciones, relaciones entre objetos, etc. Dos factores principales determinan el nivel de dificultad de cada problema específico: i) los diferentes niveles de información que se utilizan, y ii) la complejidad del modelo que se emplea con el objetivo de explicar las observaciones. La elección de una representación adecuada para los datos toma una relevancia significativa cuando se tratan invariancias, dado que estas siempre implican una reducción del los grados de libertad del sistema, i.e., el número necesario de coordenadas para la representación es menor que el empleado en la captura de datos. De este modo, la descomposición en unidades básicas y el cambio de representación dan lugar a que un problema complejo se pueda transformar en uno de manejable. Esta simplificación del problema de la estimación debe depender del mecanismo propio de combinación de estas primitivas con el fin de obtener una descripción óptima del modelo complejo global. Esta tesis muestra como los Modelos de Variables Latentes reducen dimensionalidad, que teniendo en cuenta las simetrías internas del problema, ofrecen una manera de tratar con datos parciales y dan lugar a la posibilidad de predicciones de nuevas observaciones.Las líneas de investigación de esta tesis están dirigidas al manejo de datos provinentes de múltiples fuentes. Concretamente, esta tesis presenta un conjunto de nuevos algoritmos aplicados a dos áreas diferentes dentro de la Visión por Computador: i) video análisis y sumarización y ii) datos range 3D. Ambas áreas se han enfocado a través del marco de los Modelos Generativos, donde se han empleado protocolos similares para representar datos. / The majority of problems in Computer Vision do not contain a direct relation between the stimuli provided by a general purpose sensor and its corresponding perceptual category. A complex learning task must be involved in order to provide such a connection. In fact, the basic forms of energy, and their possible combinations are a reduced number compared to the infinite possible perceptual categories corresponding to objects, actions, relations among objects... Two main factors determine the level of difficulty of a specific problem: i) The different levels of information that are employed and ii) The complexity of the model that is intended to explain the observations.The choice of an appropriate representation for the data takes a significant relevance when it comes to deal with invariances, since these usually imply that the number of intrinsic degrees offreedom in the data distribution is lower than the coordinates used to represent it. Therefore, the decomposition into basic units (model parameters) and the change of representation, make that a complex problem can be transformed into a manageable one. This simplification of the estimation problem has to rely on a proper mechanism of combination of those primitives in order to give an optimal description of the global complex model. This thesis shows how Latent Variable Models reduce dimensionality, taking into account the internal symmetries of a problem, provide a manner of dealing with missing data and make possible predicting new observations. The lines of research of this thesis are directed to the management of multiple data sources. More specifically, this thesis presents a set of new algorithms applied to two different areas in Computer Vision: i) video analysis and summarization, and ii) 3D range data. Both areas have been approached through the Generative Models framework, where similar protocols for representing data have been employed.

We present techniques for joint modeling of binomial and rank response data using the Bayesian paradigm for inference. The motivating application consists of results from a series of assessments on several primate species. Among 20 assessments representing 6 paradigms, 6 assessments are considered to produce a rank response and the remaining 14 are considered to have a binomial response. In order to model each of the 20 assessments simultaneously, we use the popular technique of data augmentation so that the observed responses are based on latent variables. The modeling uses Bayesian techniques for modeling the latent variables using random effects models. Competing models are specified in a consistent fashion which easily allows comparisons across assessments and across models. Non-local priors are readily admitted to enable more effective testing of random effects should Bayes factors be used for model comparison. The model is also extended to allow assessment-specific conditional error variances for the latent variables. Due to potential difficulties in calculating Bayes factors, discrepancy measures based on pivotal quantities are adapted to test for the presence of random effects and for the need to allow assessment-specific conditional error variances. In order to facilitate implementation, we describe in detail the joint prior distribution and a Markov chain Monte Carlo (MCMC) algorithm for posterior sampling. Results from the primate intelligence data are presented to illustrate the methodology. The results indicate substantial paradigm-specific differences between species. These differences are supported by the discrepancy measures as well as model posterior summaries. Furthermore, the results suggest that meaningful and parsimonious inferences can be made using the proposed techniques and that the discrepancy measures can effectively differentiate between necessary and unnecessary random effects. The contributions should be particularly useful when binomial and rank data are to be jointly analyzed in a parsimonious fashion.

Most psychoneuroimmunology research examines individual proteins; however, some studies have used summed score composites of all available inflammatory markers without evaluating the appropriateness of this decision. Using three different samples (MIDUS-2: N = 1,255 adults, MIDUS-R: N =863 adults, and ACE: N = 315 adolescents), this study investigates the dimensionality of eight inflammatory proteins (C-reactive protein (CRP), interleukin (IL)-6, IL-8, IL-10, tumor necrosis factor-α (TNF-α), fibrinogen, E-selectin, and intercellular adhesion molecule (ICAM)-1) and compares the resulting factor structure to a) an “a priori” factor structure in which all inflammatory proteins equally load onto a single dimension (a technique that has been used previously) and b) proteins modeled individually (i.e., no latent variable) in terms of model fit, replicability, reliability, temporal stability, and their associations with medical history and depression symptoms. A hierarchical factor structure with two first-order factors (Factor 1A: CRP, IL-6, fibrinogen; Factor 2A: TNF-α, IL-8, IL-10, ICAM-1, IL-6) and a second-order general inflammation factor was identified in MIDUS-2 and replicated in MIDUS-R and partially replicated in ACE (which unfortunately only had CRP, IL-6, IL-8, IL-10, and TNF-α but, unlike the other two, has longitudinal data). Both the empirically-identified structure and modeling proteins individually fit the data better compared to the one-dimensional “a priori” structure. Results did not clearly indicate whether the empirically-identified factor structure or the individual proteins modeled without a latent variable had superior model fit. Modeling the empirically-identified factors and individual proteins (without a latent factor) as outcomes of medical diagnoses resulted in comparable conclusions, but modeling empirically-identified factors resulted in fewer results “lost” to correction for multiple comparisons. Importantly, when the factor scores were recreated in a longitudinal dataset, none of the individual proteins, the “a priori” factor, or the empirically-identified general inflammation factor significantly predicted concurrent depression symptoms in multilevel models. However, both empirically-identified first-order factors were significantly associated with depression, in opposite directions. Measurement properties are reported for the different aggregates and individual proteins as appropriate, which can be used in the design and interpretation of future studies. These results indicate that modeling inflammation as a unidimensional construct equally associated with all available proteins does not fit the data well. Instead, empirically-supported aggregates of inflammation, or individual inflammatory markers, should be used in accordance with theory. Further, the aggregation of shared variance achieved by constructing empirically-supported aggregates might increase predictive validity compared to other modeling choices, maximizing statistical power. / Psychology

Currently most batch processes run in an open loop manner with respect to final product quality, regardless of the performance obtained. This fact, allied with the increased industrial importance of batch processes, indicates that there is a pressing need for the development and dissemination of automated batch quality control techniques that suit present industrial needs. Within this context, the main objective of the current work is to exemplify the use of empirical latent variable methods to reduce product quality variability in batch processes. These methods are also known as multiway principal component analysis (MPCA) and partial least squares (MPLS) and were originally introduced by Nomikos and MacGregor (1992, 1994, 1995a and 1995b ). Their use is tied with the concepts of statistical process control (SPC) and lead to incremental process improvements. Throughout this thesis three different sets of industrial sets of data, originating from different batch process were analyzed. The first section of this thesis (Chapter 3) demonstrates how MPCA and multi-block, multiway, partial least squares (MB-MPLS) methods can be successfully used to troubleshoot an industrial batch unit in order to identify optimal process conditions with respect to quality. Additionally, approaches to batch data laundering are proposed. The second section (Chapter 4) elaborates on the use of a MPCA model to build a single, all-encompassing, on-line monitoring scheme for the heating phase of a multi-grade batch annealing process. Additionally, this same data set is used to present a simple alignment technique for batch data when on-line monitoring is intended (Chapter 5). This technique is referred to as pre-alignment and it relies on the use of a PLS model to predict the duration of new batches. Also, various methods for dealing with matrices containing different sized observations are proposed and evaluated. Finally, the last section (Chapter 6) deals with end-point prediction of a condensation polymerization process. / Thesis / Master of Applied Science (MASc)

In this thesis we provide statistical and model-based data mining methods for pattern detection with applications to biomedical and healthcare data sets. In particular, we examine applications in costly acute or chronic disease management. In Chapter II, we consider nuclear magnetic resonance experiments in which we seek to locate and demix smooth, yet highly localized components in a noisy two-dimensional signal. By using wavelet-based methods we are able to separate components from the noisy background, as well as from other neighboring components. In Chapter III, we pilot methods for identifying profiles of patient utilization of the healthcare system from large, highly-sensitive, patient-level data. We combine model-based data mining methods with clustering analysis in order to extract longitudinal utilization profiles. We transform these profiles into simple visual displays that can inform policy decisions and quantify the potential cost savings of interventions that improve adherence to recommended care guidelines. In Chapter IV, we propose new methods integrating survival analysis models and clustering analysis to profile patient-level utilization behaviors while controlling for variations in the population’s demographic and healthcare characteristics and explaining variations in utilization due to different state-based Medicaid programs, as well as access and urbanicity measures.

Background: Frailty, a state indicating vulnerability to poor health outcomes, is a common condition in later life. However, research and intervention progress is hindered by the current lack of a consensus frailty definition and poor understanding of relationships between frailty and depression. Objectives: The goal of this research is to understand the interrelationships between frailty and depression among older adults. Specifically, this project aims 1) to examine the construct overlap between depression and three definitions of frailty (biological syndrome, medical burdens, and functional domains), 2) to determine the degree to which this overlap varies by age, gender, race/ethnicity and other individual characteristics, 3) to evaluate how the association between frailty and depression influences prediction of adverse health outcomes. Methods: This project uses data from the 2004-2012 Health and Retirement Study (HRS), an ongoing, nationally-representative cohort study of adults over the age of 55. Frailty was indexed by three alternative conceptual models: 1) biological syndrome, 2) cumulative medical burdens, and 3) functional domains. Depressive symptoms were indexed by the 8-item Center for Epidemiologic Studies Depression (CESD) scale. Latent class analysis and confirmatory factor analysis were used to assess the construct overlap between depressive symptoms and frailty. Latent growth curve modeling were used to evaluate associations between frailty and depression, and to estimate their joint influence on two adverse health outcomes: nursing home admission and falls. Results: The measurement overlap of frailty and depression was high using a categorical latent variable approach. Approximately 73% of individuals with severe depressive symptoms, and 85% of individuals with primarily somatic depressive symptoms, were categorized as concurrently frail. When modeled as continuous latent factors, each of the three frailty latent factors was significantly correlated with depression: biological syndrome (ρ = .67, p <.01); functional domains (ρ = .70, p <.01); and medical burdens (ρ = .62, p <.01). Higher latent frailty trajectories were associated with higher likelihood of experiencing nursing home admission and serious falls. This association with adverse health outcomes was attenuated after adjustment for depression as a time-varying covariate. Conclusions: Findings suggest that frailty and frailty trajectories are potentially important indicators of vulnerability to adverse health outcomes. Future investigations of frailty syndrome, however it is operationalized, should account for its substantial association with depression in order to develop more accurate measurement and effective treatment.

Latent variable models (LVMs) are commonly used in the scenario where the outcome of the main interest is an unobservable measure, associated with multiple observed surrogate outcomes, and affected by potential risk factors. This thesis develops an approach of efficient handling missing surrogate outcomes and covariates in two- and three-level latent variable models. However, corresponding statistical methodologies and computational software are lacking efficiently analyzing the LVMs given surrogate outcomes and covariates subject to missingness in the LVMs. We analyze the two-level LVMs for longitudinal data from the National Growth of Health Study where surrogate outcomes and covariates are subject to missingness at any of the levels. A conventional method for efficient handling of missing data is to reexpress the desired model as a joint distribution of variables, including the surrogate outcomes that are subject to missingness conditional on all of the covariates that are completely observable, and estimate the joint model by maximum likelihood, which is then transformed to the desired model. The joint model, however, identifies more parameters than desired, in general. The over-identified joint model produces biased estimates of LVMs so that it is most necessary to describe how to impose constraints on the joint model so that it has a one-to-one correspondence with the desired model for unbiased estimation. The constrained joint model handles missing data efficiently under the assumption of ignorable missing data and is estimated by a modified application of the expectation-maximization (EM) algorithm.